Automation Without Tears



    Recent advances in Machine Learning — Artificial Intelligence for the marketing folk — are all over the news. Dreary headlines like "A.I. Expert Says Automation Could Replace 40% of Jobs in 15 Years" have employees fearing for their jobs. Meanwhile, Wall Street analysts and businessmen (yes, they are mostly men) sharpen their pencils — and knives — as they prepare to make millions of dollars through massive job cuts.


    These existential fears and profit calculation are not only wrong but dangerous — to employees, customers, and the bottom line.


    Automation is just a means to an end. It can be used to deliver better value to customers, to save money, or both. It can also be sub-optimal, have a negative impact on consumers, and result in costly loses.


    The danger of focusing only on cost savings, is that you may end up like Oscar Wilde's cynic, who knows the price of everything but the value of nothing. After all, automation is commoditization, and commoditization is death.


    The challenge, therefore, is to use automation in ways that help your business serve customers better than the competition. Caveat emptor.



    Automating Jobs Creates Jobs


    The fear, or hope, that technology will automate away all our jobs has been a recurrent theme since the Industrial Revolution. Fortunately, the facts are as stubborn as an ox: Despite massive technological change, the unemployment rate has remained stable at around 5 percent for well over a century.


    This is all the more striking considering how the labor force grew by leaps and bounds during this period. Initially, through massive population growth, later by the entry of women into the labor force (often as computers, later as programmers, and almost always at a fraction of men's pay).



    Put differently, labor saving technology enabled the creation of millions of new jobs, Moreover, these jobs were often far safer, less physically demanding, and higher paying.


    Of course, there were also winners and losers in this process. After all, the share of the U.S. workforce employed in agriculture did fall from 41 percent in 1900, to 2 percent in 2000. But by and large the winners far outweighed the losers.



    Automating Jobs vs Automating Tasks


    To understand how labor replacing technologies can increase the demand for labor, it helps to break down the problem. Specifically, we should stop thinking of technology automating jobs wholesale, in favor of technology automating some of the tasks that make up jobs.


    That is, jobs are made out of tasks.


    For example, the job of making omelettes involves the tasks of breaking the eggs, beating them, and frying them. For some, it also involves the task of doing the dishes afterwards, though this task has been largely automated by dishwashers.


    Now consider the job of a radiologist. A key part of their job is examining images (X-rays and the like) for evidence of tumors, or other ailments. Yet, computers have become so good at this task that Geoffrey Hinton, a founding father of mother AI, quipped: "It’s quite obvious that we should stop training radiologists".


    Not so fast, retort the economists. Below is a list of the 29 tasks performed by radiologists. Of these, only 2 involve examining images. Hardly the ticket to automate away the radiologists.



    The reality is more complicated. For example, as automated image recognition becomes cheaper and better, it could be that more tests are done, and more radiologists are needed to interpret, explain, and manage the increased workflow. It could also be that nurses could manage these workflows. And so on. However, one thing seems almost certain: AI will not fully replace radiologists overnight.


    Consider also Automated Teller Machines (ATMs). These were introduced in the 1970s with the explicit aim of, you guessed it, automating tellers. So what happened? Over the next few decades the employment of tellers doubled.


    It turns out that freeing tellers from the routine job of counting cash and cashing checks, enabled banks to redeploy them as customer reps in their drive towards relationship banking. To be sure, not all tellers succeeded at making the transition from counting cash to selling mortgages, but at the very least the opportunity was there, and overall employment increased.


    Examples like these abound. For example, early spreadsheet software might have sounded the death knell of bookkeepers and accountant's. Fortunately for them, counting things, and simple arithmetic, are but small tasks in their jobs. Even today, the Bureau of Labor Statistics expects accounting jobs will gain market share over the next ten years.


    Similarly, the automation of many aspects of shoe manufacturing at the turn of the 19th century led to a rapid expansion in employment in the shoe manufacturing industry, as cheaper shoes opened up a massive market.


    As the above examples illustrate, technology does not typically automate jobs, so much as tasks within jobs. Moreover, the substitution of technology for labor tasks was seldom 1:1, but rather N:1, 1:N, 0:N, N:0, and M:N (e.g. N machine tasks to 1 labor task, 1 machine task to N labor tasks, and so on).


    For example, when the manufacture of shoes was mechanized at the turn of the 19th century, the production of shoes by hand involved 72 separate tasks, while the machine production of the same shoes involved 173 separate tasks.



    A Modern Economic Theory of Automation


    A recent paper by MIT economists Acemoglu and Restrepo argues that automation at the firm level can have be broken down into three different effects:

    • A displacement effect, as machinery or software replace labor, the demand for labor decreases.

    • A productivity effect, as the remaining labor is able to produce much more with less. This makes labor much more valuable, increasing the demand for labor in non-automated tasks. This underlies "an economic reality that is as fundamental as it is overlooked: tasks that cannot be substituted by automation are generally complemented by it." (Autor2015)

    • A reinstatement effect, as new technologies create new value-added tasks in which labor has a comparative advantage. This too increases the demand for labor. Such effects can be substantial. For example, "about half of employment growth over 1980–2015 took place in occupations in which job titles or tasks performed by workers changed." (Acemoglu and Restrepo, p. 4).

    The first effect is negative for labor, whereas the other two are positive. The net overall effect within a firm will vary on a case-by-case basis, though there is some evidence that overall negative effects may be more prominent in middle skill jobs, like data entry (Autor2015).


    Finally, you should be aware that there are also aggregate — i.e. across firms — effects (see Acemoglu and Restrepo). Examples include the displacement of self-employed artisan shoe makers by mass produced shoe manufacturers, or the 17.5 percent increase in the volume of trade after eBay introduced automated translations (which enabled more cross-border transactions).


    As a manager or employee, how can you make sure the net effect of automation is as positive as is possible for customers, investors, and employees?.



    Good and Bad AI


    Just as there is good and bad types of cholesterol, so there is good, and bad AI.


    Bad AI is the type of technology that focuses solely on displacing labor, without increasing its productivity, or creating new value added tasks. Call centers are an example of bad AI (or worse).


    Sure, such systems may field many more calls per minute than a small tyrant's bureaucracy but that is surely the wrong measure of productivity. The correct measure is the number of satisfied customers per call.


    Anecdotally, it would seem call centers are very unproductive. Except, perhaps, at the task of annoying customers right before handing them over to agents, whose task is then much more difficult, and thankless.


    By contrast, good AI displaces labor from some tasks, makes labor way more productive at the remaining tasks, and adds new high value-added tasks to the mix.


    Drug discovery, is an example where AI can help identify promising proteins faster, with greater accuracy, thus making the downstream testing carried out by humans more productive. Another example, mentioned above, is that of automated translation in eBay. See Dougherty and Wilson (2018) for many more examples.


    In their book, Dougherty and Wilson (2018) argue that the recent wave of automation has a "missing middle". By this they mean that there is too much focus on the extremes, i.e. human versus machine, and too little focus on the middle, i.e. human plus machine. Acemoglu and Restrepo (2019) echo this sentiment.


    Is bad AI outpacing good AI? Why? And what can you do about it?



    Bad Hombres


    Acemoglu and Restrepo (2019) offer a number of informed guesses as to why this is happening. To me, the most persuasive explanation has to do with herd mentality, fads, and non-economic incentives.


    My hunch is this has something to do with the way we, as a society, still promote incompetent men to leadership roles. Such men tend to see business as a tough guy, gladiatorial contest.


    In this view, firing a whole bunch of employees is just the ticket to signaling your tough guy credentials, as well as your business savvy for saving money. Unfortunately, the Bad Hombres malaise appears to be quite widespread, specially in male dominated Silicon Valley, and the startup world.


    This is a dangerous mentality. Bad AI is just simple automation. Simple automation is basically commoditization. And commoditization is death.


    Consider this simple scenario. Today, it would be easy for a company like Starbucks to automate their baristas. They could also automate their cashiers. While they are at it, they could also reduce their footprint, to account for the missing barista and a cashier. And…, and if you continue seeking efficiencies in this manner you likely end with that marvel of modern technology — the vending machine. I wonder what that would do for the stock price, if not the egos of the Bad Hombres behind the scheme.


    Fundamentally, there is a risk that the fervor for automation leads us to bypass the global optima, in favor of a local one. Tesla is a case in point. I doubt their 100 percent automation strategy makes full economic sense. Besides, the obsession with automation for automation sake may be inherently flawed. After all, such large scale automation often fails, or is plain impossible, even for the likes of Amazon.



    Is This Time Different


    The Bad Hombres theory is just a hunch. Another plausible explanation is that AI is unlike any previous technology, is mostly bad, and so past experiences are not helpful in predicting the coming dystopia.


    This requires a brief detour into the nature of AI.


    Traditional software (spreadsheets, databases, word processors, web interfaces, etc.), like the mechanization of labor, focused on well defined, programmable tasks. For example, when you enter the formula "=A1+A2" in Cell A3 of Excel, you are essentially programing the computer to add the contents in cells A1 and A2 (assuming they can be added). It is a dumb instruction.


    Now consider a more complex instruction to a robot: "Tell me if you see a chair". Early versions of AI tackled this problem using long chains of dumb instructions, like "IF it has four legs AND a seat AND a back THEN it is a chair". This proved to be a tedious, inexact, and under performant.


    Modern AI eschews dumb instruction in favor of teaching by example. Thus, instead of telling the computer what a chair is, you simply:

    1. Label thousands of photos of chairs as "Chair", and thousands of photos of other things as "Not a Chair";

    2. Run the photos and associated labels by the computer, so it can learn to predict the label from the data in the photo;

    3. Et voilà, show the computer a new photo, and it will give you a predicted label ("Chair", "Not a Chair").

    By using examples we have turned it into a prediction task, which is why AI can be thought of as a powerful prediction machine. In the first instance we tried to tell the computer, via detailed programs, what a chair was. In the second instance we simply teach it to predict labels on the basis of images.


    Throughout, there is no presumption that the computer learns what a chair is. For the most part, all we typically care about is the quality and usefulness of the predictions, not their depth.


    Back to labor and automation.


    What this means is that computers can now do tasks, like identifying a chair, or a tumor, that we cannot even begin to describe explicitly through program or rules. This is a considerable achievement, no doubt, but hardly the end of human labor as we know it.


    Prediction tasks are important tasks, to be sure, but typically they are just one task among many. Besides, only a fraction of prediction tasks have enough data, of sufficient quality, to enable performant automation.


    Maybe, the future is not so distinct from the past.



    Getting to Good AI Without Tears


    If AI is just a continuation of a long cycle of innovation extending as far back as the Industrial revolution; if it focuses mostly on automating just some tasks; and if AI is not inherently of the bad type; then there is reason to hope that AI can truly be a harbinger of a bright, and more prosperous future.


    Here are some suggestions to accomplish this:


    1. Focus on the customer, value added, and competitive advantage. AI and automation is just a means to that end.

    2. Build trust. Trust is critical to enabling fast technology adoption, and the experimentation needed to engender good AI. Overcome fears by breaking down the problem (it is about automating tasks, not jobs), and by exposing stakeholder to the limitations of AI.

    3. Beware of local optima, and excessive automation. Focus on value unlocks over cost savings, and on automating the routine, in ways that increase productivity, and enable new high-value added tasks.

    4. Avoid Bad Hombres. There is a massive opportunity cost to being stuck in poisonous politics.

    If I had to summarize it all up in a simple formula, I would say that the key to getting AI right is to focus on the value unlocks, as opposed to the cost savings. To make this transparent, you should adopt statements like


    "We are using AI to automate tasks X. Y, and Z so that we can deliver $$$ in value to our consumers",

    as opposed to statements like


    "We are using AI to automate tasks X. Y, and Z so we can save $$$".

    The former reflects a growth mindset, the latter an immiserating one.


    Your customers, investors, and employees will thank you.

    Follow me

    © 2020 by Fernando Martel García.

    Contact Me

    Contact form

    • Twitter Clean
    • White LinkedIn Icon
    • White Google+ Icon
    • GitHub-Mark-Light-32px